Our FDA comment on AI in early-phase trials
Back to all InsightsAnalysis

Everyone is racing toward the patient and stopping in the same place

Two announcements in one week, from opposite ends of the industry, both advancing on the patient and both halting at the same line. The line is not an accident, and it is exactly where the evidence says outcomes are decided.

Publication dateAugust 2026
Share to

Two things happened in the same week of August 2026, from companies with almost nothing in common.

iRhythm agreed to buy VitalConnect for $287.5 million, extending an ambulatory cardiac monitoring business into inpatient and hospital-to-home care, and picking up a biosensor platform that reads up to eleven physiological parameters.[1] And Google announced Health Guardian for Pixel Watch and Fitbit: blood pressure trends, insulin resistance estimates, and breathing emergency detection, on hardware tens of millions of people already wear.[2]

One is a medical device company buying its way into more of the stack. The other is a platform company adding physiology to a consumer product. Both are moving toward the patient. Both stop in exactly the same place, and the place they stop is more interesting than either announcement.

Google drew the line itself

Read what Health Guardian actually does, in Google's own framing. Blood pressure is delivered as one data point on the first day of each month, derived from pulse and motion rather than a cuff. Insulin resistance is a monthly score with tips attached. And the company states plainly that the features are not a substitute for diagnostic tools and do not provide clinical advice.[2]

Then the detail that says the most: breathing emergency detection, the one feature with genuine urgency attached, is launching in several European countries and not yet in the United States.[2]

None of that is a criticism. It is a company being careful and telling you precisely where it intends to stand. A monthly number is not a response loop. A trend delivered to a consumer is not a clinical finding delivered to a care team. Google built detection at a scale nobody can match, and stopped at the boundary where a signal would have to become somebody's clinical responsibility.

iRhythm stops in the same place, from the other direction

The iRhythm deal is the opposite strategy arriving at the same edge. Rather than adding physiology to consumer hardware, it consolidates clinical-grade monitoring into a single vendor's platform spanning ambulatory, inpatient and hospital-to-home settings.[1]

There is a quiet admission inside it. A cardiac monitoring company concluded that cardiac monitoring alone was not enough and paid $287.5 million for eleven-parameter capability. Multi-modal turns out to be the requirement, not the upsell.

But a vertically integrated platform solves the integration problem for one vendor's devices and leaves it exactly where it was for everyone else's. If each device maker follows the same path, a health system ends up holding several monitoring platforms that do not speak to each other, each with its own path into the chart, its own security review, its own governance conversation. The fragmentation does not disappear. It gets more expensive and better funded.

The wall is where the evidence says outcomes live

This would be an academic observation if the gap between detection and action were a minor implementation detail. It is not. It is the whole finding of the last fifteen years of remote monitoring research.

TELE-HF and BEAT-HF, together enrolling more than three thousand heart failure patients, delivered monitoring data competently and found no reduction in readmission or death.[3][4] TIM-HF2 reduced mortality, and differed from the null trials chiefly in that a clinical team monitored around the clock and adjusted therapy promptly.[5]

Put those together and the conclusion is uncomfortable for everyone shipping detection. The signal was never the scarce thing. What changes outcomes is whether the signal reaches something positioned to act, quickly, and that has always been a staffing problem rather than a sensing problem.

Which means the industry is racing hardest at the half of the problem that was already solved, and stopping at the half that decides whether any of it matters.

Why almost nobody crosses it

Not for lack of ambition. The reasons are structural, and they differ by who you are.

A device maker cannot cross it without building for competitors' hardware. The moment iRhythm's platform ingests a rival's sensor, the platform stops being a reason to buy iRhythm's devices. Vertical integration and neutrality are mutually exclusive, and the acquisition picks a side.

A platform company cannot cross it without becoming regulated healthcare infrastructure: business associate agreements with health systems, PHI custody, audit obligations, and liability for what an algorithm tells a patient to do. Google's own disclaimer and its decision to withhold the emergency feature from the US market are that calculation, made visible.

And the remote monitoring platforms have not crossed it either, though they are closest. Across that layer the pattern is consistent: AI arrives as a clinician-facing copilot. Documentation, summarization, risk stratification, worklist prioritization. Useful, and pointed at the staff rather than the patient. Even the most AI-forward player in the category keeps a human on the patient-facing clinical endpoint, with nurse practitioners approving medication changes and clinicians conducting the conversations.

Each of those is a rational decision. Together they leave a gap that widens every time somebody ships better detection.

The nearest exception makes the point rather than breaking it. A small number of agent companies have crossed to the patient-facing side, and the best funded of them now publishes its escalation rates with a major academic health system on its reference list. But look at how they crossed: by becoming the agent vendor, shipping one model under one safety case, evaluated before deployment. That is a different structural position from a layer any agent can run inside, and it relocates the safety question rather than settling it. Pre-deployment evaluation certifies the model that was tested. It says nothing, at the moment of sending, about the message that is about to send.

What crossing it actually requires

It is worth being concrete, because "close the loop" is the kind of phrase that sounds like a product and is actually a list of unglamorous obligations.

A signal has to arrive with its provenance intact, so that a value can still be defended a year later. Anything acting on it has to run somewhere a hospital's security review will accept. Whatever it produces for a patient has to be checked before it is sent, against rules a clinically accountable person wrote, rather than reviewed afterwards once it has already arrived. When something is held back, that has to reach a human who can act, with the original attached. The result has to land in the record system the team already works in. And the whole chain has to be reconstructable, because a governance committee will eventually ask.

None of that is intelligent. All of it is required, and none of it is what a device maker or a platform company is trying to build.

What we think, and what we are not claiming

We build in that gap, so treat this as an interested reading rather than a neutral one.

What we will not claim is that closing the loop has been proven to change outcomes. It has not, not with an automated responder. The null trials tell us data alone does not work. TIM-HF2 tells us a staffed loop does. There is now a preprint suggesting an agent can perform the triage step at a level comparable to individual clinicians on retrospective readings, which is genuinely new and is still a component test rather than an outcomes trial.[6] The honest position is that automating the response loop is a reasonable thing to test prospectively, not a demonstrated result.

What does seem clear is where the industry's attention is, and where it is not. Detection is being industrialized by companies with enormous advantages in it. The thing that made TIM-HF2 work is being industrialized by almost nobody, because it requires an accountability posture rather than a technical one.

That is a strange allocation of effort, and the two announcements this week made it unusually easy to see.

Sources

  1. iRhythm Technologies, "iRhythm to Acquire VitalConnect, Expanding Its Cardiac Monitoring Platform Across Ambulatory, Inpatient and Hospital-to-Home Care," August 2026 - globenewswire.com
  2. Google, "Track subtle changes in your body with Health Guardian features on Pixel and Fitbit," August 2026 - blog.google; feature and availability detail via mobihealthnews.com
  3. Chaudhry SI, et al. (TELE-HF), "Telemonitoring in Patients with Heart Failure," NEJM, 2010 - nejm.org
  4. Ong MK, et al. (BEAT-HF), "Effectiveness of Remote Patient Monitoring After Discharge of Hospitalized Patients With Heart Failure," JAMA Internal Medicine, 2016 - jamanetwork.com
  5. Koehler F, et al. (TIM-HF2), "Efficacy of telemedical interventional management in patients with heart failure," The Lancet, 2018 - thelancet.com
  6. Kim S, Kung TH, Verma H, et al., "From Days to Minutes: An Autonomous AI Agent Achieves Reliable Clinical Triage in Remote Patient Monitoring," arXiv preprint 2603.09052, March 2026 (not peer reviewed) - arxiv.org

Talk through where this is going.

We'll walk through your program and where the regulatory and market shifts leave it.

Moving forward together